Simplified artificial neural network based online adaptive control scheme for nonlinear systems.

Complex neural network structures may not be appropriate for adaptive control applications, as large number of variable adaptive parameters have to be updated in each time step, thus increasing the computational burden, which makes the real-time implementation of the controller difficult. In this study, the ability of the neural networks to approximate non-linear dynamic systems is used to derive a neuro-based adaptive control with minimalistic architecture. A linear neural identifier is devised, which emulates a local linear model of the system by online adjustment of its parameters. Stability and optimal rate of convergence is ensured through an adaptive learning rate, determined using Lyapunov stability theorems. The novelty of the control scheme lies in its minimalistic neural structure comprising of a single-linear neuron and, therefore, does not impose excessive computational burden on the system, making it feasible for real-time application. To assert our claims, benchmark examples from different domains are used to illustrate the effectiveness of the proposed controller. The neuro-controller is used in a water-lift plant to control the height of the water in a storage tank. Another example of a moving cart holding an inverted pendulum, an inherently unstable system that forms the basis of the robot-control mechanism, is also used. The controller is also tested on a complex non-linear higher-order power system to enhance stability by effectively damping the electromechanical oscillations. The superior performance of the controller is demonstrated by comparing with other recently reported controllers. Additional advantages of the proposed scheme include model-free control and requirement of only local measurements. The proposed method has potential applications in control problems that require adaptability, computational simplicity, and quick response. [ABSTRACT FROM AUTHOR]